Has GPT-6-Astra reached embodied intelligence?

Problem

GPT-6-Astra, a general-purpose model, has started solving robot manipulation and navigation tasks that used to be dominated by task-specialized models such as VLAs. There is a need to lay out systematically what this model can actually do and what it means for research directions.

Method

The analysis has four parts. (1) The model was evaluated with a unified protocol on 12 benchmarks covering tabletop manipulation, humanoids, mobile robots, dexterous hands and navigation. The protocol allows in-context learning (ICL) using a fixed "knowledge package" built from training-set demonstrations, and extends execution steps by up to 10x. (2) The authors compared harness designs that combine a high-level "System 2" reasoner (GPT-6-Astra) with a low-level "System 1" controller (π0.5), as well as action representations and action chunking. (3) Results were validated on 8 real robots. (4) The authors propose a prototype next-generation benchmark in which tasks come from real demand, no domain demonstrations are given, and time keeps passing while the model reasons.

Results

• The authors report 81.9% on RoboDojo and 87.3% on DexJoCo under ICL with an extended budget (10x), and above 95% on the LIBERO family and SIMPLE, close to saturation.
• On 4 shared SIMPLE tasks, task-level ICL raised the success rate under the standard budget from 70.8% (zero-shot) to 90.8%.
• Raising the execution budget 10x roughly tripled the success rate on RoboDojo and more than doubled it on Bench2Dex. On the long-horizon navigation benchmark LHPR-VLN, however, it rose only slightly.
• Closed-loop collaboration, where GPT-6-Astra reviews actions proposed by a VLA, raised performance substantially but also had a high token cost. End-effector commands worked better than direct joint-angle control.

Limitations

The headline numbers come from conditions that differ from the official protocols (ICL based on training-set demonstrations, up to 10x execution steps), so they are hard to compare directly with existing SOTA. The authors also note that ICL does not transfer well beyond the range of the demonstrations, that inference cost and execution time are unrealistic for commercial deployment, and that verifying physical success, recovering from irreversible failures and tracking long-horizon progress are weak points. In real robot settings, raising success rates by adding steps does not work because of the cost of collecting demonstrations and safety constraints.

Why it matters

As general-purpose frontier models saturate much of the robot benchmark landscape, this paper points out the gap between benchmark scores and real deployment capability. It is useful for laying out what evaluations should measure going forward, such as cost, time constraints and evaluation without demonstrations.

Key terms

ICL(In-context Learning)

A way of solving a task by drawing on examples or material placed in the input context, without changing the model weights. In this paper, a knowledge package built by analyzing training demonstrations is consulted at test time.

System 1 / System 2

An architecture that splits work between a fast low-level controller (System 1, e.g. π0.5) and a slower high-level model that reasons in depth (System 2), and has them work together.

Action Chunking

A method that outputs actions in chunks of several steps rather than one step at a time. Adjusting the chunk length to the situation can reduce the number of model calls.

Original

• arXiv abstract
• PDF
• HTML full text
• Code
• Project
• Hugging Face

Authors: SenseNova, Joint Research Team, :, Zhongang Cai, Ziang Cao, Tailai Chen, Han Chong Chua, Linjun Dai, et al. · Posted 2026-10-15 · License arXiv 비독점 배포 라이선스 · Based on the full text

This review is an AI-written summary of the paper. Please treat the original paper as the reference.